2 x 1080 Ti in NON-SLI = 22GB of VRAM?

[phongle123]

Hello, I understand that in SLI of 2x1080 Ti I would not get the benefit of the 2nd 11GB of VRAM and would be capped at 11GB still in SLI.

 

My end result is not for gaming. Rather it is to render using Nvidia iRay which takes full advantage of VRAM with scenes that contain large textures or just multiple objects that fill up the GPUs VRAM.

 

Right now I have an 8GB and a 4GB Nvidia card. When I use it to render an iRay Scene it uses 7/8GB and 3/4GB of each card respectively. So, it seems that these cards are independent of each other as they are separate cards. Since the 8GB card isn't also stuck at 3/8GB when the 4GB card is using 3/4GB. I'm not sure whether or not I'm actually getting the 8+4=12 GB of VRAM in this case.

 

So if I used 2 x1080 Ti to render independently (NOT in SLI; but still using both GPUs to render). Am I correct to assume I have 22 GB of VRAM for the render?

[Cereal5]

I don't know your answer, but if you're in the US and your 8GB card happens to be a 1080, wanna sell it to me?

[KarathKasun]

Not really.  Are you rendering stills or animation?

 

For still images the scene is usually split into chunks for each device doing computation, and all devices need the full set of assets.  Devices with less ram will get a smaller chunk or will end up streaming data from system ram.

 

In some cases with animation the devices will render one frame on each device with the same problem as above, cards with less ram will probably perform significantly slower because they need to stream data over the bus.

[phongle123]

Just now, KarathKasun said:

Not really.  Are you rendering stills or animation?

Stills. But give me the answer to both.

[Droidbot]

it would seem so - i'd still try it with another card and see if rendering performance improves.

 

theoretically applications like this would use vram as texture cache only during the raytracing process.

[KarathKasun]

2 minutes ago, phongle123 said:

Stills. But give me the answer to both.

Edited the answer in, figured I should have already posted it.

[phongle123]

4 minutes ago, Droidbot said:

it would seem so - i'd still try it with another card and see if rendering performance improves.

 

theoretically applications like this would use vram as texture cache only during the raytracing process.

The actual results seems like it does since I'm actually maxing out the VRAM on both cards and that the 8GB card is NOT being pulled down by the 4GB card when I look at the VRAM usage in Afterburner.

Though, sometimes on larger scenes my 4GB card crashes and only my 8GB card is being used. 

 

But for the rendering improvement part. Yes when using the 2nd card (the 4GB card) real-time iRay rendering is very much improved and I can actually work better with it using the real-time iRay mode instead of just the textured mode. The 8GB card alone was unworkable when using the real-time iRay.

[phongle123]

12 minutes ago, KarathKasun said:

and all devices need the full set of assets

Does this mean that if they were NOT in SLI. It would take longer for the scene to prepare since the textures would have to be put into the VRAM of both GPUs.

 

Whereas in SLI, since it's working as one. It would take half the time since the textures are only being pushed into 1 GPU to prepare the scene?

[KarathKasun]

4 minutes ago, phongle123 said:

Does this mean that if they were NOT in SLI. It would take longer for the scene to prepare since the textures would have to be put into the VRAM of both GPUs.

 

Whereas in SLI, since it's working as one. It would take half the time since the textures are only being pushed into 1 GPU?

No matter how you set them up both compute nodes will have a full set of assets for the scene in ram.  Also, no matter how they are set up, assets are loaded from the CPU and they do not transfer directly from GPU to GPU.

 

SLI does NOT make the GPU's function as one, it only adds the functionality to copy the output image to a single card for displaying both outputs on a single output.